Fine-grained 서베이 논문 3편

1. A Survey of Recent Advances in CNN-Based Fine-Grained Visual Categorization
2. A Systematic Evaluation: Fine-Grained CNN vs. Traditional CNN Classifiers
3. Deep Learning for Fine-Grained Image Analysis: A Survey

3번째 논문.

• Fine-grained의 특징(도전)
  • inter-class similarity, intra-class variability
• 데이터셋 정리
  • 유명한거 순서
    • CUB200-2011 > Standford Car > FGVC Aircraft > Oxford Flowers > Standford Dog >...
• 푸는 방법 2가지
  • Strongly-supervised
    • Category말고도 어느 위치에 있는지 box or location이 추가적으로 존재함.(이미지마다)
  • Weakly-supervised - 3가지
    • Category만 존재
    • 3가지 방향
      • Part-based Approach
        • 중요한 Part를 찾으려고 노력하는 architecture를 만듦
      • End-to-End
        • Architecture 변형 접근
        • Loss 접근
    • with External information
      • Web image / knowledge graph / text가 존재.

End-to-End 자세히

1. Architecture
   1. high-order mixing of CNN features가 주임.
   2. Hierarchical Bilinear Pooling(HBP-ECCV 2018)
   3. Deep bilinear transformation(DBT-NIPS 2019)
2. Loss
   1. GCE loss(AAAI 2020) : focus on top-k confusing negative classes
   2. hinge loss(API-Net AAAI 2020) : score ranking regularization

Fine-Grained CNN vs Traditional CNN

• 결론
  • Traditional CNN 승
    • DenseNet161최고
    • Traditional CNN이 ImageNet pretraining에서 이득을 봄.

3번째 논문 : Image classification말고도 여러가지 있음

• Few-shot learning 논문 소개
  • Piecewise classifier mappings IEEE TIPS 2019
• 더 읽어야 함. 좋은게 많음.